Zero-to-CAD 1M matters because it treats CAD generation as a program synthesis problem rather than only geometry regression: instead of meshes or labeled feature maps, you get executable CadQuery scripts that encode parametric construction history. That makes samples immediately replayable, editable, and directly useful for models that must produce interpretable CAD programs or editable 3D assets.
What Sets It Apart
- Million-scale, executable sequences: ~999,633 validated CadQuery programs (train/val/test splits) that run to produce solids—so models can be trained to emit runnable code rather than opaque geometry. This is unusual at this scale in CAD datasets.
- Agentic synthesis pipeline: an LLM iterates with an execution/validation loop (execute_and_validate, documentation lookup, grep) to repair and converge on valid constructions, yielding realistic operation distributions (fillets, lofts, sweeps, booleans, patterns) and named parameters useful for program-conditioned generation.
- Rich multimodal artifacts: each item includes 8 rendered views, STL and STEP exports, cadquery ops metadata, and precomputed DINOv3 visual embeddings plus a FAISS IVF-PQ index for nearest-neighbor search—so you can train/image-index, do retrieval-augmented generation, or evaluate image→CAD pipelines without building your own render/embedding stack.
Who It's For & Trade-offs
Great fit if you need large-scale, replayable parametric CAD data for training generative models (text→CAD, image→CAD, code-conditioned CAD), building retrieval systems over CAD assets, or studying program-repair/agentic synthesis workflows. Look elsewhere if you require real-world CAD provenance, industry-grade units/standards, or manufacturing-ready assemblies—Zero-to-CAD is fully synthetic, constrained to scale conventions (max units) and reflects LLM priors rather than measured production frequencies. Also beware the dataset size: use streaming access or selective download to avoid heavy local storage and memory costs.
Where It Fits
Compared to sketch-extrude corpora, Zero-to-CAD emphasizes a broader operation vocabulary and executability at scale. Versus real CAD repositories, it sacrifices real-world provenance for coverage, validation consistency, and program interpretability—valuable for model training but less suited as a drop-in production CAD library.
Methodological Note
The dataset was generated by placing an LLM in a feedback loop with a CadQuery executor and documentation retrieval tools; generations were geometrically validated (topology, contiguity, complexity) and filtered. Precomputed visual embeddings and neighbor tables enable fast similarity search without re-rendering or re-embedding.
